The processing a model spends thinking about a request before it answers, and the main reason two runs of the same prompt can cost very different amounts.
Also known as inference-time compute, thinking budget, reasoning effort
For most of the last decade, a model’s capability was fixed at training time. Test-time compute shifts part of that budget to the moment of use: the model generates intermediate reasoning, weighs alternatives, checks its own work, or samples several candidate answers and picks the best one. More thinking generally means better answers on hard problems, paid for in time and money.
Providers now expose this as a dial, whether it is called reasoning effort, a thinking budget, or simply a choice between a fast model and a deliberate one. The tokens a model produces while thinking are billed even though nobody ever reads them. That is the detail that surprises people the first time a scaled workflow’s invoice arrives.
Test-time compute is where AI stops behaving like a flat subscription and starts behaving a little like media spend.
It changes per-asset economics. A prompt that costs a fraction of a cent on a fast model can cost twenty times that at a high reasoning setting. Across one headline, irrelevant. Across six thousand product descriptions, it is a line item somebody has to own.
It is a quality dial you can set per task. Trimming copy to a character count needs no deliberation. Sorting four hundred open-ended survey responses into a usable segmentation does. Running both at the same setting means overpaying for one and underserving the other.
It shapes what you can promise on timing. Deliberate models are slower by design. A real-time personalization idea that looks effortless in a demo can fall apart in production when each response takes eleven seconds and the page has already been abandoned.
A studio builds an automation that drafts localized ad copy for fourteen markets. In testing, on a handful of examples at the highest reasoning setting, the output is excellent and the cost is invisible. Rolled out across a full seasonal catalog, the same workflow runs nine thousand times and the bill becomes something the account lead has to explain in a status meeting. The version that survives splits the job: a fast, cheap model handles the mechanical translation and length-fitting, and a high-reasoning pass runs only on the two hundred headline lines where nuance actually decides whether the ad works.
The automations and agents module of the workshop teaches you how to build AI workflows that compress the busywork without taking the craft out of the studio.